A woman reviews a futuristic dashboard of AI agents, identity checks, security alerts, and automated workflows in an office.
Microsoft used Ignite 2026 to make a large bet on identity. AI agents are piling up in enterprise environments, and the company wants every one of them to have an identity, an owner, a risk score and a retirement date, the same way people do. On November 17, Yina Arenas, Microsoft's Corporate Vice President and Chief Product Officer for Identity and Network Access, published a long list of Microsoft Entra announcements on the Microsoft Entra Blog. She grouped them under three goals: strengthen the identity foundation, secure agent access at scale, and turn customer identity into a competitive advantage.

Most of the list is in preview, and some of the generally available labels don't yet match Microsoft's own documentation. Here's what was announced, what Microsoft Learn adds, and what Windows and Microsoft 365 admins should do first.

A woman reviews a futuristic dashboard of AI agents, identity checks, security alerts, and automated workflows in an office. The pitch: agent sprawl is the new shadow IT​

Microsoft's argument is simple. Agents inherit the identity problems organizations already have, including excess privileges, standing access and weak governance, and then multiply them. In the blog's words, today's identity backlog becomes tomorrow's agent security problem. The post cites industry projections of roughly 1.3 billion agents in production by 2028. That's a forecast quoted in a vendor announcement, not a measured count.

The more concrete claim is from Microsoft's own telemetry: on Azure AI workloads, tool and agent abuse made up 12% of observed attack activity. The post gives no methodology or denominator, so read it as an indicator, not a benchmark.

Independent coverage picked up the same theme. Redmondmag described the Entra announcement as one that introduces additional controls for managing agents as distinct identities rather than relying exclusively on the permissions of human users. That separation between the human and the agent working for them runs through nearly every announcement below.

Section summary: Microsoft is treating agents as first-class identities. The supporting statistics are vendor-sourced and best read as directional.

Identity foundations: passkeys, verification and less standing access​

Phishing-resistant sign-in​

According to Arenas, passkeys are now the default for new Microsoft Entra accounts. The post cites a 2026 FIDO Alliance report finding that 57% of organizations still rely on phishable methods such as SMS codes. Those methods are exposed to adversary-in-the-middle phishing during sign-in, onboarding and account recovery.

  • High assurance credentials (preview): pairs passkeys with third-party identity verification at onboarding, before access to sensitive resources or meetings, and during account recovery. Named partners are AU10TIX, IDEMIA, CLEAR, 1Kosmos, Entrust, Persona and TrueCredential. The idea is that a passkey is only as trustworthy as the process that issued it.
  • Secure enterprise browser for BYOD (preview): Edge for Business plus Entra Global Secure Access. Users sign in to an Edge work profile, even on a personal device, to reach private apps. A Secure Web and AI gateway covers their browsing. Microsoft presents this as a narrower alternative to legacy VPN access.

Privileged Identity Management expands​

The PIM preview extends just-in-time access to more users, resources, application roles and AI agents. Highlights:

  • Users can activate roles and connect to Azure and Azure Arc VMs from the My Access portal.
  • Admins can monitor and manage privileged sessions and review session recordings.
  • AI-based recommendations suggest least-privileged roles.
  • A new PIM-only mode enforces role-assignment policies.
  • Bulk management helps cut standing access at scale.

Conditional Access policy plans: what Microsoft Learn adds​

Policy plans (preview) let admins define shared assignments, resources and conditions once, then add individual enforcement rules underneath. Templates include ones for AI agents. Microsoft Learn's concept page, last updated November 9, fills in details the blog leaves out:

  • Plans don't replace existing policies. Existing Conditional Access policies keep working unchanged. Both legacy policies and plan rules are evaluated at sign-in, and access requires every applicable policy and rule to be satisfied, with no precedence between them.
  • Limits: each plan needs at least one rule and can hold up to 100. Each tenant can have up to 100 plans.
  • Exclusions only add up. A rule can't narrow the plan's targeting or remove a plan-level exclusion, but it can add exclusions of its own. Emergency-access exclusions set at the plan level can't be undone by an individual rule.
  • One identity type per plan level. Shared conditions can't target users and agents together. To cover both in one plan, leave identity targeting to the individual rules.
  • "All resources" only at rule level. This assignment isn't available in shared conditions.
  • Preview caveat on location: during preview, plans that target users with a Location condition issue one-hour access tokens without an instant IP enforcement check. If you rely on named locations as a hard boundary, test this before you depend on it.
  • Role required: creating, changing or deleting a plan requires the Conditional Access Administrator role.

Each rule appears separately in sign-in logs, which makes troubleshooting easier than unpicking one huge policy.

Groups, guests and tenants​

  • Dynamic membership groups (preview): MemberOf Next combines membership from selected groups, and Transitive Reports To extends membership across a whole reporting hierarchy.
  • Guest lifecycle policies (GA): Entra ID Governance evaluates guest sponsorship, inactivity and invitation status, with reporting that shows which guests are affected before and during enforcement. Consultant Gethyn Ellis's Day 1 roundup also lists this as generally available: Entra ID Governance can now review guest sponsorship, inactivity and invitation status against policies.
  • Tenant Governance (preview): cross-tenant configuration management and resource policy to deploy, monitor and enforce configurations and detect drift across tenants. Tenant Governance also extends to Agent 365 and Copilot in a separate multitenant preview.
  • Backup and Recovery Premium (preview): 30-day retention and integration with Microsoft 365 Backup. A companion hard deletion protection preview guards key objects, including Agent IDs, from permanent deletion. Ellis adds that UK organisations should include agent identities and Conditional Access in recovery exercises. That advice applies everywhere.
  • Microsoft MCP Server for Enterprise (GA): connects AI tools and agents to data in Entra, Purview and Intune so they can answer queries and handle admin tasks.

Section summary: passkeys by default, wider just-in-time access, and Conditional Access plans that add a new layer on top of existing policy. The plans have documented limits and a location caveat that admins need to plan around.

Securing agents: discover, identify, govern, contain​

Discovery and identity​

The headline preview is Entra Agent ID for discovered local agents. Using Defender and Global Secure Access, it can detect 60 local agent types, including Claude Code, OpenAI Codex and GitHub Copilot CLI, and give each one an Agent ID. Microsoft's Ignite Book of News describes the rollout as starting with GitHub Copilot CLI and Claude Code, to separate agent from human activity and help secure their access. Treat the "60 types" figure as the eventual scope, with coverage starting with those two tools.

A second preview gives digital teammates their own agent user account and ID. Conditional Access, ID Protection and governance can then apply to the agent's work separately from the person it works for.

This builds on earlier foundations. Microsoft Learn's release notes show General Availability - Microsoft Entra Agent ID platform in April 2026, using standard protocols such as OAuth 2.0, MCP, and A2A. Merill Fernando's Entra.news also notes that since July 2026 every new Copilot Studio agent gets an Entra Agent ID, and you can no longer opt out. Ignite extends coverage beyond the cloud to the developer laptop.

Lifecycle governance​

  • Dynamic permissions for permission remediation (preview): finds agent identities with excessive privileges in connected business apps and suggests fixes.
  • Access packages for digital teammates (preview): approval and time-limit policies, joining existing access packages for Microsoft and third-party cloud agents.
  • Automated lifecycle management of agent identities (announced as GA): policies check whether an agent has a human sponsor, is still needed and has been active recently. If not, Entra notifies people and then disables or deletes the agent. Virtualization Review repeated the GA status and the claim that a default policy disables identities that have been inactive for 60 days.

The fine print on that 60-day policy​

The current Microsoft Learn page for the Agent ID default inactivity policy, last updated October 28, describes a more gradual process than "disable at 60 days":

Days inactiveDefault action
60Notify the agent's sponsors and owners
90Disable the agent identity
455Delete the agent identity

Other operational details from the same page:

  • No license is required for the default inactivity policy. Custom lifecycle policies with extra rules require appropriate licensing.
  • Default timings can't be changed. Admins can set the scope to all Agent IDs or to all Agent IDs except selected ones, and add fallback notification recipients (group mailboxes are supported) for agents with no sponsor or owner.
  • Only owners or admins can recover a disabled or deleted agent. Sponsors can't re-enable or restore one.
  • Custom policies take precedence. The default policy always applies last.
  • Supported roles: Lifecycle Workflows Administrator, AI Administrator or Agent ID Administrator.

That page also still labels Agent ID lifecycle policies as preview, with no SLA and a warning against production use. The Ignite post calls automated lifecycle management generally available, and the documentation may simply not have caught up yet. Until it does, check what your tenant's admin center shows before you treat the feature as production-supported.

How to switch on the default inactivity policy​

Microsoft Learn documents these steps:

  • Sign in to the Microsoft Entra admin center with one of the roles listed above.
  • Go to ID Governance > Lifecycle Workflows > Lifecycle Policies > Agent ID (Preview).
  • Select Default Agent Identity Inactivity Lifecycle Policy.
  • Open the Policy Configuration tab and toggle Enable policy.
  • Review the configuration and select Save.

What success looks like: Entra periodically evaluates in-scope agents, and those inactive for 60 days enter the notify-then-disable process. The most common mistake: seasonal or rarely used agents, such as a year-end reporting bot, get disabled. Use the All Agent IDs except selected scope to exclude them before you enable the policy, not after finance comes looking for you. Learn also notes that a user can bring an agent back to active status by interacting with it, so nobody can quietly keep an idle agent alive without using it.

Risk-adaptive and network controls​

  • Dynamic Agent Risk (Agent 365 preview): combines signals from identity, data, endpoints, cloud, network, runtime and the agent's own behavior. Examples include whether it runs in a container or loose in the OS, and whether the data it's after is sensitive. Signals come from Entra, Defender and Purview, with more planned.
  • Conditional Access for assistive agents (preview): separates a person from the local agent acting for them, so a policy can block a risky agent while the employee keeps working. A new actions and destinations option extends protection from whole apps to individual sensitive operations.
  • Conditional Access for Windows 365 for Agents (GA): Conditional Access now covers the Cloud PCs your agents use.
  • Agentic traffic acquisition and agent-type-aware network controls (preview): Global Secure Access visibility and policy for agent traffic. Microsoft says these work alongside an existing SASE setup, but it hasn't documented compatibility limits.
  • MCP firewall (preview): policies covering MCP servers, tools, resources, methods, protocol versions and connection security. Microsoft's BRK343 demo shows a policy that lets Claude Code and Claude Desktop use only trusted servers and tools.
  • Network-layer DLP with Purview and Global Secure Access (GA): covers human and on-behalf-of agent traffic, including unsanctioned AI apps.

Windows itself is part of this. Microsoft's Copilot Ignite post says Microsoft is strengthening security and containment for local agents running on devices through deeper integration with Intune, Defender, and Windows with Microsoft Execution Containers (MXC). If your developers run coding agents on Windows 11 workstations, these controls are aimed directly at them.

Section summary: Microsoft's agent controls follow a lifecycle: discover, assign identity, scope access, watch risk, retire. The parts that are generally available today are lifecycle policy (with the documentation caveat above), Conditional Access for Windows 365 for Agents and network DLP. Most of the rest is still preview.

Customer identity: External ID gets more flexible​

The Entra External ID announcements are all previews and aimed at developers:

  • HTML templates with JavaScript for branded sign-up and sign-in. Microsoft says these are rolling out now, with full availability by December 7.
  • Bring your own UX authentication: self-hosted authentication pages, while External ID still enforces access controls and handles single sign-on.
  • Multiple first-factor methods: passkey, one-time passcode, or email and password.
  • Native Authentication API v2 and SDKs for mobile and single-page apps, plus an AI agent toolkit that analyzes a codebase to onboard apps or migrate from v1 to v2.
  • Self-service credential management APIs: passkey registration and password updates through low-privilege delegated permissions, with email and phone APIs to follow.
  • TOTP for external users: a second factor using codes from Microsoft Authenticator or a supported third-party authenticator app.

Austrian motorway operator ASFINAG appears as a reference customer serving 4 million users. That's a vendor-selected testimonial, not independent performance evidence.

Analysis: what's solid and what isn't yet​

Microsoft has a clear and well-structured agent-security story, and it's notably vendor-neutral about which agents it governs, with Claude Code and Codex treated the same as GitHub Copilot CLI. Still, a few points deserve some skepticism:

  • Preview is not production. Most of these features are preview, and Microsoft's preview terms mean no SLA and possibly incomplete capabilities.
  • The GA labels don't always match the docs. The lifecycle policy mismatch is a good reason to check the admin center and Microsoft Learn before you rely on any status label.
  • Licensing is still unclear. The blog post gives no SKU details for most features. Apart from the free default inactivity policy, confirm entitlements before you plan anything around them.
  • Signals depend on the Microsoft stack. Dynamic Agent Risk and local agent discovery work best if you already run Defender, Purview and Global Secure Access. Shops using a different security vendor will get less out of them.

The practitioner line Microsoft quoted is the right one to end on: if you're not doing it for humans today, you'll struggle with agents. Organizations that already enforce just-in-time admin access, clean up guests and keep Conditional Access tidy will find these tools fit right in. Those that don't will now have to fix the same gaps for agents as well as people.

Practical next steps for admins​

  • Inventory first. Find out which agents already have Agent IDs, especially Copilot Studio agents, since new ones get them automatically.
  • Enable the default inactivity policy carefully. Exclude seasonal agents and set fallback recipients before you switch it on.
  • Pilot policy plans in a test tenant or on a narrow scope. Remember that plans add to existing policies, and check the location-condition token behavior.
  • Add agent identities and Conditional Access to recovery drills, especially if you evaluate Backup and Recovery Premium.
  • Watch the developer endpoints. Local coding agents are the first discovery targets, so coordinate with whoever owns your Defender and Intune configuration.

The BRK343 demo session, with Arenas, John Savill and Nichole Peterson, is scheduled for Wednesday, November 18, at 10:15 AM PT. Microsoft has also scheduled an "Ask Microsoft Entra anything" AMA for December 2 and a webinar series from December 15 to 17.


Correction (November 17, 2026): This article originally named the product "Microsoft MCP Server for IT Admins." The correct product name is Microsoft MCP Server for Enterprise, as listed in Microsoft's November 17 Entra Ignite blog. The generally available status and description were correct.

 

References

  1. AI is changing access. See how Entra helps you stay ahead. Microsoft Entra Blog 2026-11-17T16:00:00+00:00
  2. Microsoft Entra news announcements from Microsoft Ignite November 2026 techcommunity.microsoft.com
  3. What are Microsoft Entra Conditional Access policy plans? (preview) - Microsoft Entra ID | Microsoft Learn learn.microsoft.com